作者: Genung L Clapper
DOI:
关键词:
摘要: 966, 211. Automatic speech recognition. INTERNATIONAL BUSINESS MACHINES CORPORATION. Dec. 19, 1962 [Dec. 21, 1961], No. 47865/62. Heading G4R. A complex waveform, e.g. a signal is analysed into series of discrete digital samples by detecting the presence different selected components there being means responsive to changes in any take all components. The'speech from microphone 10, Fig. 1, applied pre-amplifier 12 which compressor purpose improve signal-to-noise ratio. An automatic gain control obtained integrating input signal, effects compression and also passed on line 59 other circuits described below. The compressed 11 pre-emphasis 20-22 amplify broad bands pass signals frequency selector 27-32. Amplifier 18 non-selective feeds sibilant noise detector. 24 responds only low "voice" frequencies no circuit provided this channel. All channels are integrator-shaped 42- 55 each output pulses when energy associated above certain threshold value. outputs except that relating voice seven matrix drivers 88 comprising diode gate transistor amplifier. 48 train representing their fundamental speaker. pulse integrated pair integrator arranged opposite side middle point such way if normal remains at zero volts, as for input. If below positive generated 645 above, negative signal. rising therefore gives falling vice versa. Circuit 91 detects or giving leads 667 669 inflections. constant produced both. To avoid inflection absence voicing, these gated with "voice present" two connected 92. measure This an intensity digitiser 62 analogue converter combinational 61, 63 represents range instant roughness 60 differentiators produce long short excursions respectively, together so excursion followed causes integrator, indicates quality "roughness" passes twelve far bi-polar transient detectors 64 differentiator consisting bridge diodes upward downward Any change channel produces corresponding sampling values existing store 44. These delayed 74 92 then present inputs ring counter 86 steps select turn columns 44 successive enter columns. first sample 80 generates representation logarithm time elapsing after beginning word. For capacitor -12 volt source converted form using devices above. remaining entered times. Provision made prevent intervals less than 12. 5 milliseconds 42-55 unless received minimum amplitude duration. read-out 99 read out data required. shown greater detail 2 (not shown) Analogue-to-digital converters: transistors 563, 567 2h biased voltages potential divider 563 conducts -3 volts give via 569 load 573. At -6 causing 571 conduct, lead 575 operating 565 bias -9 it switches off, thereby removing again now both 573 575.